The short version
Each time a tool does the research and drafting behind a piece of work, the person whose name is on it misses the effort that would have taught them the subject, and the shortfall builds up. A 2025 study from the MIT Media Lab called what accumulates cognitive debt. The study measured people writing essays, not lawyers giving advice, and it should be read that way; the name fits a problem supervisors already know.
What it is
The trainee who can send the research note but cannot answer the partner's first question about it. The associate whose draft is fluent and whose reasons for it are not their own. The output is there; the understanding that would normally have been built while producing it is not.
The professional position does not move because a tool did the drafting. In R (Ayinde) v London Borough of Haringey and Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin), judgment of 6 June 2025, the Divisional Court compared a lawyer's reliance on AI research to reliance "on the work of a trainee solicitor or a pupil barrister" (paragraph 8): the duty to check stays with the lawyer who uses it (National Archives). The Bar Council's guidance says AI tools "are not substitutes for a barrister's independent legal research verification, analysis, and judgment" (Considerations when using ChatGPT and generative artificial intelligence software based on large language models, last reviewed 25 November 2025, paragraph 24).
A sighting
The study is Nataliya Kosmyna and others, Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task (arXiv, first posted 10 June 2025). Fifty-four participants wrote essays in three sessions, in one of three groups: using a language model, using a search engine, or using nothing. Eighteen came back for a fourth session, in which the language-model group wrote unaided and the unaided group used the language model. The researchers recorded brain activity with EEG and analysed the essays.
What it reported, in its own words:
- "LLM users displayed the weakest connectivity", against the strongest for those writing unaided.
- "Self-reported ownership of essays was the lowest in the LLM group and the highest in the Brain-only group."
- "LLM users also struggled to accurately quote their own work." The project page puts it more plainly: the group "fell behind in their ability to quote from the essays they wrote just minutes prior" (brainonllm.com).
What it did not measure matters as much. The task was essay writing, not legal analysis. The authors say the findings are "focused on writing an essay in an educational setting and may not generalize across tasks", list a "limited number of participants recruited from a specific geographical area" among the limitations, and the project page notes that, as of June 2025, the paper "has not yet been peer-reviewed, thus all the conclusions are to be treated with caution and as preliminary". The study ran over four months; it says nothing about skill over a career, or about legal work.
The example below is made up to show the pattern. A trainee produces a note on whether a notice was validly served, citing three authorities and reaching a clear conclusion. The supervising solicitor asks which of the three authorities carries the conclusion, and why the other two are there. The trainee cannot say. The ranking, and the reasoning behind it, were the tool's.
Why it happens
Producing the work is how a lawyer usually comes to understand it: reading the cases, discarding the ones that do not fit, writing the reasons down. When a tool does that part, the output arrives without the understanding that the process would have built. The study's abstract records that "Cognitive activity scaled down in relation to external tool use."
What to say back
"Which authority does the conclusion rest on, and why are the others there? Which steps in the reasoning did you work through yourself, and which did the tool supply?"
Origin
- 6 June 2025. Ayinde and Al-Haroun: the Divisional Court compares reliance on AI research to reliance on a trainee's work (paragraph 8).
- 10 June 2025. Kosmyna and others post the study on arXiv, with "cognitive debt" in its title.
- 25 November 2025. The Bar Council's reviewed guidance: AI tools are not substitutes for a barrister's own analysis and judgment (paragraph 24).
- Now. The study is still about essays. Whether the same debt builds up in legal training has not been measured in any source cited here.